summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorDavid S. Miller <davem@davemloft.net>2016-03-18 22:25:46 -0400
committerDavid S. Miller <davem@davemloft.net>2016-03-18 22:25:46 -0400
commit7fa7728d23e7928dfb6e9f5540a97fb2d562b989 (patch)
tree4b57cd88bdb39ff9025cafb3f9b8fcbe22660b6c
parent79d3b59a93ba25f3b2c72eb4099c189d41d30204 (diff)
parenta3382e408b645b4f68ec01f9c048e356c62598fb (diff)
Merge branch 'rds-buffer-tuning'
Sowmini Varadhan says: ==================== RDS: TCP: tunable socket buffer parameters Patch 1 uses sysctl to create tunable socket buffer size parameters. Patch 2 removes an unuused constant. v2: use sysctl v3: review comments from Santosh Shilimkar, Eric Dumazet v4: review comments from Hannes Sowa ==================== Signed-off-by: David S. Miller <davem@davemloft.net>
-rw-r--r--net/rds/tcp.c145
1 files changed, 134 insertions, 11 deletions
diff --git a/net/rds/tcp.c b/net/rds/tcp.c
index ad60299b088b..61ed2a8764ba 100644
--- a/net/rds/tcp.c
+++ b/net/rds/tcp.c
@@ -52,7 +52,34 @@ static LIST_HEAD(rds_tcp_conn_list);
static struct kmem_cache *rds_tcp_conn_slab;
-#define RDS_TCP_DEFAULT_BUFSIZE (128 * 1024)
+static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
+ void __user *buffer, size_t *lenp,
+ loff_t *fpos);
+
+int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
+int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
+
+static struct ctl_table rds_tcp_sysctl_table[] = {
+#define RDS_TCP_SNDBUF 0
+ {
+ .procname = "rds_tcp_sndbuf",
+ /* data is per-net pointer */
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = rds_tcp_skbuf_handler,
+ .extra1 = &rds_tcp_min_sndbuf,
+ },
+#define RDS_TCP_RCVBUF 1
+ {
+ .procname = "rds_tcp_rcvbuf",
+ /* data is per-net pointer */
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = rds_tcp_skbuf_handler,
+ .extra1 = &rds_tcp_min_rcvbuf,
+ },
+ { }
+};
/* doing it this way avoids calling tcp_sk() */
void rds_tcp_nonagle(struct socket *sock)
@@ -66,15 +93,6 @@ void rds_tcp_nonagle(struct socket *sock)
set_fs(oldfs);
}
-/* All module specific customizations to the RDS-TCP socket should be done in
- * rds_tcp_tune() and applied after socket creation. In general these
- * customizations should be tunable via module_param()
- */
-void rds_tcp_tune(struct socket *sock)
-{
- rds_tcp_nonagle(sock);
-}
-
u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
{
return tcp_sk(tc->t_sock->sk)->snd_nxt;
@@ -272,8 +290,34 @@ static int rds_tcp_netid;
struct rds_tcp_net {
struct socket *rds_tcp_listen_sock;
struct work_struct rds_tcp_accept_w;
+ struct ctl_table_header *rds_tcp_sysctl;
+ struct ctl_table *ctl_table;
+ int sndbuf_size;
+ int rcvbuf_size;
};
+/* All module specific customizations to the RDS-TCP socket should be done in
+ * rds_tcp_tune() and applied after socket creation.
+ */
+void rds_tcp_tune(struct socket *sock)
+{
+ struct sock *sk = sock->sk;
+ struct net *net = sock_net(sk);
+ struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
+
+ rds_tcp_nonagle(sock);
+ lock_sock(sk);
+ if (rtn->sndbuf_size > 0) {
+ sk->sk_sndbuf = rtn->sndbuf_size;
+ sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
+ }
+ if (rtn->rcvbuf_size > 0) {
+ sk->sk_sndbuf = rtn->rcvbuf_size;
+ sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
+ }
+ release_sock(sk);
+}
+
static void rds_tcp_accept_worker(struct work_struct *work)
{
struct rds_tcp_net *rtn = container_of(work,
@@ -295,20 +339,60 @@ void rds_tcp_accept_work(struct sock *sk)
static __net_init int rds_tcp_init_net(struct net *net)
{
struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
+ struct ctl_table *tbl;
+ int err = 0;
+
+ memset(rtn, 0, sizeof(*rtn));
+ /* {snd, rcv}buf_size default to 0, which implies we let the
+ * stack pick the value, and permit auto-tuning of buffer size.
+ */
+ if (net == &init_net) {
+ tbl = rds_tcp_sysctl_table;
+ } else {
+ tbl = kmemdup(rds_tcp_sysctl_table,
+ sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
+ if (!tbl) {
+ pr_warn("could not set allocate syctl table\n");
+ return -ENOMEM;
+ }
+ rtn->ctl_table = tbl;
+ }
+ tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
+ tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
+ rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
+ if (!rtn->rds_tcp_sysctl) {
+ pr_warn("could not register sysctl\n");
+ err = -ENOMEM;
+ goto fail;
+ }
rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
if (!rtn->rds_tcp_listen_sock) {
pr_warn("could not set up listen sock\n");
- return -EAFNOSUPPORT;
+ unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
+ rtn->rds_tcp_sysctl = NULL;
+ err = -EAFNOSUPPORT;
+ goto fail;
}
INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
return 0;
+
+fail:
+ if (net != &init_net)
+ kfree(tbl);
+ return err;
}
static void __net_exit rds_tcp_exit_net(struct net *net)
{
struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
+ if (rtn->rds_tcp_sysctl)
+ unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
+
+ if (net != &init_net && rtn->ctl_table)
+ kfree(rtn->ctl_table);
+
/* If rds_tcp_exit_net() is called as a result of netns deletion,
* the rds_tcp_kill_sock() device notifier would already have cleaned
* up the listen socket, thus there is no work to do in this function.
@@ -383,6 +467,45 @@ static struct notifier_block rds_tcp_dev_notifier = {
.priority = -10, /* must be called after other network notifiers */
};
+/* when sysctl is used to modify some kernel socket parameters,this
+ * function resets the RDS connections in that netns so that we can
+ * restart with new parameters. The assumption is that such reset
+ * events are few and far-between.
+ */
+static void rds_tcp_sysctl_reset(struct net *net)
+{
+ struct rds_tcp_connection *tc, *_tc;
+
+ spin_lock_irq(&rds_tcp_conn_lock);
+ list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
+ struct net *c_net = read_pnet(&tc->conn->c_net);
+
+ if (net != c_net || !tc->t_sock)
+ continue;
+
+ rds_conn_drop(tc->conn); /* reconnect with new parameters */
+ }
+ spin_unlock_irq(&rds_tcp_conn_lock);
+}
+
+static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
+ void __user *buffer, size_t *lenp,
+ loff_t *fpos)
+{
+ struct net *net = current->nsproxy->net_ns;
+ int err;
+
+ err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
+ if (err < 0) {
+ pr_warn("Invalid input. Must be >= %d\n",
+ *(int *)(ctl->extra1));
+ return err;
+ }
+ if (write)
+ rds_tcp_sysctl_reset(net);
+ return 0;
+}
+
static void rds_tcp_exit(void)
{
rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);